Professional Documents
Culture Documents
ECON2209
Slides 01
BF-01
About BF
Staff
Lecturer: Dr Minxian Yang, ASB452, 93853353, Fri 10-1
Tutors: see Tutorial Contacts
Assessment
BF-01
Class participation = 3%
2 Class-tests = 20 % (10% each)
1 Project = 17% (due week 12)
Final exam = 60%
my, School of Economics, UNSW
About BF
Objective of BF
Foster skills for analysing time series data and
capabilities of applying models, principles, techniques
in a business environment
Coverage
Focus mainly on uni-variate time series models, which
is the foundation for complex models.
Will touch upon multi-variate time series models.
About BF
Course resources
Course website:
announcements, course outline, lecture slides, tutorial
questions/answers, assignment, data, Eviews code
Library open (close) reserve
Discuss course material in consultation (not by email)
Ch.1 Introduction
Introduction
Forecast
A statement about the future values of a variable,
based on current knowledge
Ch.1 Introduction
Essence of forecasting
e.g. Try to predict the next number
a) {2, 4, ?}.
b) {2, 4, 8, ?}.
c) {2, 4, 8, 14, ?}.
How did we predict?
Inspect data; (gather information)
Find a pattern; (fit a model)
Predict according to the pattern. (extrapolate model)
BF-01
Ch.1 Introduction
Essence of forecasting
Predict by exploiting the pattern in data
data
prediction
Better data
More info about pattern
Better model
Better prediction.
Economists are better
at predicting the past than the future.
- Joseph E. Stiglitz
BF-01
Ch.1 Introduction
Uncertainty
Consider USD/AUD exchange rate.
Ch.1 Introduction
Uncertainty
Forecasting under uncertainty is challenging:
Many possible outcomes (randomness): Perfect prediction is
impossible. Forecast errors must be allowed.
Sources of uncertainty:
Outcomes are affected by human (re)actions,
Also by many small factors that are hard to pin down.
Ch.1 Introduction
Uncertainty
BF is about finding patters in data and using patters to
extrapolate:
If these patters persist, then the variables will behave with
these patters in the future.
10
Ch.1 Introduction
Forecasting Model
A good model captures major patterns in data.
eg.
Sales = Regular Sales + Disturbance
forecast object,
with past observations
pattern,
which is summarised by a model
unpredictable,
with distributional regularities
BF-01
11
Ch.1 Introduction
Topics to be covered
Framework of business forecasting
Forecast environment, loss function, info set, horizon,
parsimony principle
Statistical techniques
Time series description, classical decomposition, trend and
seasonality, ARMA models, estimation/testing, VAR models
Application emphasised
Data description (statistical/graphical), model selection,
interpretation of results, forecast statement
Implementation of models, EViews
End of Ch.1
BF-01
12
13
14
Forecast object
This is usually a future event of interest.
Forecast outcome: quantitative/qualitative
eg. Winner of the next election (qualitative)
Next quarters GDP growth rate (quantitative)
Forecast timing
eg. When will the next election be held?
When will GDP growth rate become negative?
BF-01
15
Forecast horizon
This is the number of periods to the target date.
h-Step ahead forecast
eg. 4-quarter-ahead forecast of GDP growth rate:
0.6% (quarterly)
f.o.
q1
0.7
q2
0.9
q3
0.8
q4
0.6
16
Information set
This includes historical data relevant to forecast objects.
The quality of forecasts depends on
The quality and quantity of data
The skills and experience in exploiting data
Notation
Let y be the variable of interest. The info set at t is
t = { y1 , y2 ,..., yt ; x1 , x2 ..., xt },
Everything
useful & observed
at Date t
BF-01
17
Loss function
It measures the badness of forecasts.
Notation
y:
y :
variable to be forecast;
forecast for y ;
e = y y : forecast error;
L( y, y ) : loss function, usually L = L( y y ) = L(e).
BF-01
18
Loss function
Required properties of L(e)
L(0) = 0;
(no loss with perfect prediction)
L(e) is increasing in |e|;
(further away from zero the error, larger the loss)
L(e) is continuous;
(small change in error leads to small change in loss)
L(e)
Asymmetry
under-forecast may
incur more costs.
BF-01
e
my, School of Economics, UNSW
19
Loss function
Symmetric loss functions
e.g.
Quadratic: L(e) = e2
20
Loss function
Expected loss
We average L using the distribution of y:
Expectation:
average weighted
by probability
E{L(e)} =
(1-.5)2(.5) + (0-.5)2(.5)
Optimal forecast
Choose y to minimise expected L for a given info set.
BF-01
21
Loss function
Mean squared forecast error
(quadratic loss)
MSFE = E{( y y ) 2 }
We can only
choose y-hat.
= E{( y y ) 2 } + E{( y y ) 2 }
where y = E{ y}.
E should be evaluated
relative to given info set, ie,
conditional expectation.
22
Forecast statement
This is the presentation of forecasts,
which depends on the decision environment.
Point forecast (simplest)
It is a single-value prediction.
eg. 4th quarter GNP growth point forecast = 0.6%
Prob(future y is in
Interval forecast
the interval) = 0.9
It is an interval with a coverage probability.
eg. 4th quarter GDP growth 90%-interval-forecast = [0.4%, 0.8%]
23
Parsimony principle
BF-01
800
600
400
GDP ($billion)
eg.
Data
y = Australian annual GDP
Model-1:
y = a + bYear + error
Model-2:
y = a + bYear + cYear2 + error
1000
1200
1970
1980
1990
2000
2010
Year
24
Parsimony principle
Simple model
It emphasises important features and ignores many
unimportant details.
Easy to implement
Experience: simple models often do well in forecasting.
Complex model
It captures more details of data. But details may blur
important features.
More details could contain more noises.
Harder to implement, carries more estimation errors.
forecast error
BF-01
description error
estimation error
my, School of Economics, UNSW
25
Parsimony principle
Parsimony principle
Other things being equal, simpler models are preferred.
eg.
yT +1 = yT ;
(ii)
yT +1 = yT + xT .
Unless Model (ii) has proven better, we will use Model (i)
BF-01
26
1000
800
'000
eg.
Predict
at Feb/82
Predict
at Feb/88
Predict
at Feb/94
600
400
200
0
Feb-98
Feb-96
Feb-94
Feb-92
Feb-90
Feb-88
Feb-86
Feb-84
Feb-82
Feb-80
Feb-78
27
Summary
forecast object
forecast horizon
information set
loss function
forecast statement
Parsimony principle
Forecasts are based on available information set
BF-01
28